Qualitative discrimination of coal aerosols by using the statistical evaluation of laser-induced breakdown spectroscopy data

Abstract

The present work demonstrates the capabilities of laser-induced breakdown spectroscopy (LIBS) in the qualitative analysis of carbonaceous aerosols. The model aerosols studied were generated by laser ablation in a nitrogen flow from five commercial coal samples (lignite, anthracite, Pécs-vasas brown coal, Polish brown coal, Czech brown coal) and contained sub-micron particles in a concentration exceeding 106 cm−3. Features of the LIBS spectra of these aerosol samples were characterized and it is showed that the particle detection frequency (expressed as the number of particle hits referenced to the total number of laser shots delivered) correlates with the mass concentration of the aerosol. The detection limit for coal aerosols was also estimated and found to be about 600 pg·mm−3, meaning that for detectability with the present experimental system, either the diameter of individual particles should be over 2.3 μm or their number concentration has to be large enough to exceed the above mass concentration detection limit.The possibilities for coal classification based on the statistical evaluation of the LIBS spectra of their aerosols was also investigated in detail in the laboratory. Simple comparative functions (overlapping integrals, sum of squared deviations, linear correlation) were found not to be efficient in the discrimination, even when facilitated by spectral masking, but multivariate methods (classification tree, linear and quadratic discrimination analysis) gave significantly better results. The effect of data normalization and data compression (by the multivariate curve resolution alternating least squares methodology) prior to modeling was also tested and it was found that their influence on the classification accuracy is not always positive, if any. The best performance was showed by the classification tree method (without data compression), which had a good overall accuracy of 87.2%. The validation of the model was assessed by calculating the repeatability of the classification accuracy from 15 repetitions using randomly selected subsets of the spectra. This calculation gave a repeatability value of 2.2%, which shows that the model is quite robust.